Models of molecular evolution

Models of molecular evolution have several components that each consider different aspects of the evolutionary process. The models that have been used to model protein evolution in studies of the deep phylogeny have largely differed according to three components: the exchangeability matrix \(E\), the rate of evolution, and the state equilibrium frequencies \(\Pi\). A detailed description of each component is in Supplementary text. Briefly, (1) the exchangeability matrix \(E\) describes the rate at which one amino acid changes to another. Exchangeability matrices have been used in the studies under consideration here include: Poisson (or F81), WAG, LG, GTR, as they have been used regularly in phylogenomic analyse; (2) While the exchangeability matrix describes the relative rate of different changes between amino acids, the actual rate can be further scaled, such as site homogeneous rates and Gamma-rate heterogeneity; (3) the vector of equilibrium frequencies \(\Pi\) describes the stationary frequency of amino acids, such as empirical site-homogeneous, estimated site-homogeneous, CAT site-heterogeneous³, and C10-C60 models (C models). (4) data-recoding method recode 20 amino acids are recoded into six groups based on function to account for both compositional heterogeneity and substitution saturation. Several recoding strategies have been proposed, including Dayhoff 6-state recoding, S&R 6-state recoding, KGB 6-state recoding based on exchangeability matrix of models⁴.

Models can be assembled by selecting different options for all these different components. The models that are applied in practice area heavily influenced by engineering and computational costs, as well as convention. For example, on the questions considered here, Poisson and GTR exchangeability matrices have only been used in combination with CAT site-heterogeneous models of equilibrium frequency. LG and WAG exchangeability matrices have only been used with site homogeneous estimates of equilibrium frequency. This is further confused by the abbreviations that are used for models. Papers often discuss CAT and WAG models as if they are exclusive, but these particular terms apply to non-exclusive model components– CAT refers to variation across sites and WAG a particular exchangeability matrix. CAT is generally shorthand for F81+CAT and WAG is shorthand for WAG+homogeneous equilibrium frequency estimation. One could, though, run a WAG+CAT model. To avoid confusion on this point, we always specify the exchangeability matrix first, followed by modifiers that describe the accommodation of heterogeneity in equilibrium frequencies (e.g., CAT) or rate (e.g., Gamma). If there are no modifiers, then it is implied that site homogeneous models are used.

Gene sampling

High-throughput sequencing is leading to exciting new data matrices for phylogenomic analyses, potentially including hundreds or thousands of genes to explore the root of metazoan phylogeny (Table 1). Although the number of sampled genes might be different across studies, it is reasonable to expect that a conserved core set of orthology genes (OGs) should be present across most animal genomes (e.g., animal BUSCO genes⁵). Surprisingly, gene composition and matrix overlap has been rarely assessed across studies and most studies relied on their own bioinformatic pipeline for determination of OGs (Supplementary text) for animal phylogeny.

Outgroup and ingroup taxon sampling

Outgroup selection (the taxa used to root the ingroup taxa) can potentially affect the phylogenetic inference of the root position of animal phylogeny (Fig XXOverview.C). Several studies have progressively removed more distantly related outgroup taxa to test the effect of outgroup selection to ingroup topology^???,6. Three progressively more inclusive clades have often been investigated, including Choanomalia (Metazoa plus most closely related Choanoflagellata), Holoza (Choanimalia plus more distantly related holozoans) and Opisthokonta (Holozoa + Fungi). It has been suggested the inclusion of more distant related (e.g. Fungi) outgroup taxa tend to recover Ctenophora-sister, whereas high support of Porifera-sister often recovered when more closely related outgroup taxa were used with site-heterogeneous CAT models⁶. In contrast of outgroup selection, sensitivity to ingroup sampling has received less attention than sensitivity to outgroup sampling. This may be because results have tended to be more sensitive to outgroup sampling.

Data-recoding

Compositional heterogeneity (lineage-specific differences in amino acid frequencies) was long thought has a major effect in resolving the animal tree of life, especially for the deep animal phylogeny⁷. Feuda et al. (2017)⁴ has proposed to use data-recoding method to reduce compositional heterogeneity in data matrices. By discarding information on which specific amino acid is found at each site in each species and instead focusing on which groups those amino acids belong to, they sought to reduce the impact of differences between species in which particular amino acids within those groups are most frequent. By reanalyzing Chang et al. 2015⁸ and Whelan et al. 2015⁹, they report that posterior predictive analyses indicate recoded analyses have better model adequacy than non-recoding analyses, and that Porifera-sister was strongly favored under all recoding strategies using GTR-CAT models⁴. However, how the recoding methods impact phylogenetic inference has not been well explored. Moreover, a recent simulation study has suggested that non-recoding approaches significantly outperformed recoding strategies¹⁰.

Matrix taxon composition

Taxa sampling plays a critical role in phylogenetic inference, especially for the taxa whose phylogenetic position that needs to be explored. However, several data matrices have relatively skewed taxon sampling, with more than 50% taxa sampled in bilaterians compared to other major animal lineages (Fig XXTaxon_composition). These skewed data matrices include two studies constructed from ESTs (Dunn2008² and Hejnol2009¹¹) and one study from whole-genome data (Borrowiec2015¹² ). Importantly, several earlier studies only include less than five ctenophore species. The rest data matrices have relatively balanced taxon composition, although the total number of sampled taxon varies a lot. In general, increasing taxon sampling is broadly accepted to improve phylogenetic inference and aid in the placement of animal root position. Thus, it should be noted that Whelan2017¹³ (has the most Ctenophore taxa) and Simion2017¹⁴ (contains the most taxon sampling) have much broader taxon sampling than other data matrices (Table1, Fi. XXTaxon_composition).

Fig XXTaxon_composition. Each of the primary matrices considered here, color coded by taxon sampling. Horizontal size is proportional to the number of genes sampled, vertical size to the number of taxa sampled.

Matrix gene composition and overlap

Gene sampling plays an essential role in phylogenomic analyses, and most often rely on the concatenation of sampled genes for deep animal phylogeny. It has been suggested that the concatenation assumes that phylogenetic signal from genes that do not share the species phylogeny will be overwhelmed by the signal from the majority of genes whose genealogy mirrors that of the species evolutionary history¹⁵. However, large phylogenomic data matrices potentially contain genes with conflicting or weak phylogenetic signal that can lead to incongruence phylogenomic results, especially when analyzing deeper node represents ancient divergence in the tree¹⁶. Thus, it is important to evaluate the gene composition and matrix overlap to explore level of congruence of gene sampling across different matrices. It should be noted that early studies relied on EST via Sanger sequencing contains large missing data compared to other ones (e.g., Dunn2008 and Hejnol2009).

We first used the Metazoa BUSCO dataset to evaluate gene composition from all data matrices and the result is unexpected. The animal BUSCO dataset contains 978 single-copy OGs (identified from 64 animal taxa) that are most likely shared by most animal genomes, and are commonly used to assess the completeness of genome or transcriptomic assemblies⁵. Moreover, the BUSCO dataset also are frequently used as markers in many large-scale phylogenomic studies (e.g., yeasts¹⁷, spiders¹⁸ ). Thus, it is reasonable to expect that these conserved genes should present at a higher percentage than non-BUSCO genes in most matrices. However, to our surprise, we found that gene partitions from all matrices comprised less than 50% of BUSCO genes, even with several data matrices contain less than 20 percent (Fig. XXBUSCO_annotations).

Previous reanalyses have mainly focused on outgroup selection and/or model fits in different data matrices, and the matrix overlap across different studies has been rarely investigated. Our initial intent of comparing matrix overlap across matrices was to construct a new matrix with shared taxon and gene sampling to test specific hypotheses about differences in support. The hypothesis is that although the bioinformatics pipeline used for OG determination is different, most larger data matrices should share a core set of OGs. However, our results show that surprisingly little overlap across all the data matrices from different studies (Fig XXAlignment overlap). Thus, these results suggested that gene sampling largely varies across studies, indicating that different regions of the genomes are sampled with little overlap among different data matrices.

## <ggproto object: Class CoordFixed, CoordCartesian, Coord, gg>
##     aspect: function
##     backtransform_range: function
##     clip: on
##     default: FALSE
##     distance: function
##     expand: TRUE
##     is_free: function
##     is_linear: function
##     labels: function
##     limits: list
##     modify_scales: function
##     range: function
##     ratio: 1
##     render_axis_h: function
##     render_axis_v: function
##     render_bg: function
##     render_fg: function
##     setup_data: function
##     setup_layout: function
##     setup_panel_params: function
##     setup_params: function
##     transform: function
##     super:  <ggproto object: Class CoordFixed, CoordCartesian, Coord, gg>

Fig XXBUSCO_annotations. BUSCO annotation and representation in each data matrix. (A). The number of partitions with BUSCO annotations in each matrix, relative to the number of partitions. (B). The percentage BUSCO annotations in each matrix.

Fig XXGene_composition. Each of the primary matrices considered here, color coded by the types of genes sampled (XX Ribosomal proteins, etc). Horizontal size is proportional to the number of genes sampled, vertical size to the number of taxa sampled.

Fig XXAlignment overlap. Pairwise overlap between each of the primary matrices considered here. Horizontal size is proportional to the number of genes sampled, vertical size to the number of taxa sampled. The horizontal intersection shows the proportions of shared genes, the vertical intersection shows the proportions of shared taxa.

Support for Porifera-sister and Ctenophora-sister

Here we summarized all the 131 previous phylogenetic analyses from 15 studies (Fig. support_published_analyses and Table perviouly_published_analyses). Among these 15 studies, the main conclusions of five studies strongly in favor of Porifera-sister and 10 supports Ctenophora-sister (TablexxPAPERS). Moreover, three of them are studies of reanalysis without generating any new data matrices. We found that all past phylogenetic analyses strongly support of Ctenophora-sister once site-homogeneous models were used, no matter what outgroup choice, phylogenetic inference and whether data-recoding are used. The support for Porifera-sister increases with the exclusion of more distantly related outgroups with the use of site-heterogeneious CAT (Poisson+CAT+G or GTR+CAT+G) models. It should be noted that many BI analyses with CAT model are not well resolved (either not reach convergence or do not provide significant support for the key node of the root of animal phylogeny) mainly due to computational limits of CAT models employed in PhyloBayes. The strong support of Porifera-sister can only be recovered from analyses using Poisson+CAT+G and recoding methods with GTR+CAT+G models in several matrices. Although GTR+CAT model generally has better model adequacy than Poisson+CAT or site-homogeneous models from previous studies^4,6, Porifera-sister is not always supported since several analyses strongly in favor Ctenophora-sister even with only Choanofalgellates are used as outgroups (Fig. support_published_analyses).

Fig support_published_analyses. A total of 131 analyses were transcribed from the literature (Table perviouly_published_analyses.tsv).

Comparison of iqtree and phylobayes results

Site heterogeneity in equilibrium frequency has been a major concern in tests of Ctenophora-sister and Porifera-sister. This has been addressed with CAT models. IQtree provides a new family of C models that also address site heterogeneity. Given the extensive computational cost and concerns about overparameterization of CAT models¹⁹, we compared iqtree C results to CAT results for a subset of matrices to see if they give consistent results. This would be of technical interest because it would reduce the cost of accommodating compositional heterogeneity in future analyses. For computational efficiency, we first only applied Poisson+CAT models in every matrix with Choanoflagellata as outgroup. It should be noted that most Phylobayes runs were converged, although several large matrices have not reached convergence after at least a month’s computational time.

Interestingly, we found strong supports for both hypotheses when using Poisson-CAT model in Choanimalia matrices. Similar to ML analyses, we found no strong support for either hypothesis in Moroz2013_3d matrix. Consistent with previous study, we recovered strong support of Porifera-sister in matrices of Philippe2008, Ryan2013, Nosenko2013 and Whelan2015. Interestingly, we also found a strong support of Porifera-sister in two Whelan2017 matrices and a weaker support of subsampled matrix of Simion2017. Ctenophora-sister were strongly supported in all other matrices. For the matrices have conflicted results between Poisson-CAT and C models, we further ran Poisson+CAT models with inclusion of more distantly related outgroups to evaluate how outgroup choice may affect phylogenetic inference under CAT model. We found strong support of Porifera-sister in holozoa matrices and Ctenophora-sister in all analyzed Opisthokonta matrices, indicating that the support of Porifera-sister is increasing with inclusive of the more closely related outgroups. Thus, our results show that both model choice and outgroup sampling play a major role in incongruent results of the root position of animal phylogeny in several key data matrices.

Phylogenetic signal

The distribution of phylogenetic signal over the root position of animal phylogeny has never been quantified with site-heterogeneous models. To further explore the effect of model selection and outgroup choices in the root position of animal phylogeny, we next quantify the support of phylogenetic signal over two alternative hypotheses (T1: Ctenophora-sister (Fig. 1A); T2: Porifera-sister (Fig. 1B)) to three key data matrices with different outgroup choice and models in both ML and BI framework (Fig XXPhylogenetic signal). By calculating gene-wise log-likelihood scores between T1 and T2 for every gene (\(GLS) or site (\Delta\)SLS) in each matrix, we found that Ctenophora-sister had the higher proportions of supporting genes in every analysis when using site-homogeneous models (LG or WAG in both IQtree or PhyloBayes). Moreover, the outgroup choice has little impact on the distribution of the support of phylogenetic signal in site-homogeneous models. This finding is largely consistent with the past study that majority of phylogenetic signal is strongly favored in Ctenophora-sister hypothesis with site-homogeneous models in other data matrices²⁰.

In contrast, the distribution of phylogenetic signal has never been quantified in site-heterogeneous models and compared with site-homogeneous models. Although Ctenophora-sister also had the higher proportions of supporting genes with C models, we found that the phylogenetic signal decreases in many genes using C models compared to site-homogeneous models, especially in two matrices from Ryan_2013 that nearly 30% of genes changed from strong to week signal ($lnL<2) (Fig. phylogenetic_signal B). It is not clear why the phylogenetic signal decreases in many genes with a better fitting model as suggested by model testing in IQtree, but our results indicated the C models lack of phylogenetic signal compared to site-homogeneous models in the node of root position of animal phylogeny. In contrast to the C models, the phylogenetic signal largely increased in Poisson-CAT model and more importantly, the outgroup choice has a major effect of the distribution of phylogenetic signal. We found that support for Ctenophora-sister largely decreases and Porifera-sister increases (~ 20% of sampled gene partitions) when exclusive of more distantly related outgroups with the Poisson-CAT models. The results here are further corroborated by our phylogenomic analyses in BI, which the support of Porifera-sister increases in Holozoa and Choanimalia matrices.

We also examined the level of congruence of genes that are supported in each hypothesis across different analyses in Whelan2017 matrices. Consistent with the summary of $GLS analysis, We found gene population that supports in either hypothesis are largely congruent between site-homogenous and C models, whereas little overlap was observed in Poisson-CAT models (especially with different outgroup choice). In other words. our results strongly indicated that the distribution of phylogenetic signal of the root of animal phylogeny is largely sensitive to model selection (site-homogeneous vs. site-heterogeneous models). More importantly, the site-heterogeneous CAT model is also more sensitive to the outgroup choice than site-homogeneous and C models in these key data matrices. Thus, it is clear that recovering Porifera-sister requires more constrains, including more stringent model selection, outgroup choice, and even with gene sampling.

Fig XXPhylogenetic signal. The distribution of phylogenetic signal for two alternative topological hypotheses on the earliest-branching animal lineage. Proportion of genes supporting each of two alternative hypotheses for each of ### Bias of data-recoding There is growing interest in data-recoding and previous study⁴ hoped that data-recoding would reduce potential artifacts due to differences across species in amino acid frequencies. They report that posterior predictive (PP) analyses²¹ indicate 6-state recoded analyses have better model adequacy than 20-state amino acid analyses, and “Porifera-sister was favored under all recoding strategies” in Whelan2015_D20 and Chang2015 data matrices. Here we focus on two aspects of Feuda et al. First, we found that many of their recoded analyses are actually unresolved (i.e., without strong support for either Porifera-sister or Ctenophora-sister), and that the analyses with the best posterior predictive scores do not provide strong support for Porifera-sister (Fig XXRecoidng_summary). Second, we created four new random recoding schemes by shuffling the amino acids in the SR-6 scheme (see methods). When we applied each of these randomized codes to the Whelan matrix and analyzed them under the CAT-GTR+G model with phylobayes-mpi, we observed similar results as for the empirical SR-6 recoding. Like SR-6 recoding, random recoding increases support for Porifera-sister and improves the apparent adequacy of models to explain heterogeneity of states across taxa (PP taxon hetero mean and max, Figure XXRandom_recodingB). Thus, our analyses show the impact of recoding is largely due to discarding information, not accommodating variation in amino acid composition. These findings indicate that it is premature to accept Porifera-sister and reject Ctenophora-sister using data-recoding methods and recoding can be a problematic method for addressing compositional variation (see more detailed discussion in supplementary text).

Data selection and wrangling

We retreived matrices from each publication (Table XX), storing the raw data in this manuscript’s version control repository. We manually edited some minor formatting to make the batch processing of the matrices en masse, e.g. standardizing the formatting of charset blocks. All changes made are tracked with git.

Matrix comparison and annotation

Phylogenetic analyses

MPKE AY IDGQ TRC WNLF SVH
EIFT WL QVPG NKM SCHA RYD
LCQS GK WPTI VRD YEFN MAH
IWQV TY KDLM ESH APCF GRN

Dunn et al. 2008

Dunn et al.² added Expressed Sequence Tag (EST) data for 29 animals. It was the first phylogenomic analysis that included ctenophores, and therefore that could test the relationships of both Ctenophora and Porifera to the rest of animals. It was also the first phylogenetic analysis to recover Ctenophora as the sister group to all other animals.

The data matrix was constructed using a semi-automated approach. Genes were translated into proteins, promiscuous domains were masked, all gene sequences from all species were compared to each other with blastp, genes were clustered based on this similarity with TribeMCL³², and these clusters were filtered to remove those with poor taxon sampling and high rates of lineage-specific duplications. Gene trees were then constructed, and in clades of sequences all from the same species all but one sequence were removed (these groups are often due to assembly errors). The remaining gene trees with more than one sequence for any taxon were then manually inspected. If strongly supported deep nodes indicative of paralogy were found, the entire gene was discarded. If the duplications for a a small number of taxa were unresolved, all genes from those taxa were excluded. Genes were then realigned and sites were filtered with Gblocks³³, resulting in a 77 taxon matrix. Some taxa in this matrix were quite unstable, which obscured other strongly-supported relationships. Unstable taxa were identified with leaf stability indices³⁴, as implemented in phyutility³⁵, and removed from the matrix. This resulted in the 64-taxon matrix that is the focus of most of their analyses. Phylogenetic analyses were conducted under the F81+CAT model in phylobayes³, and under the WAG model in MrBayes³⁶ and RAxML³⁷.

Regarding the recovery of Ctenophora-sister, the authors concluded:

The placement of ctenophores (comb jellies) as the sister group to all other sampled metazoans is strongly supported in all our analyses. This result, which has not been postulated before, should be viewed as provisional until more data are considered from placozoans and additional sponges.

Note that there was, in fact, an exception to strong support. An analysis of the 40 ribosomal proteins in the matrix recovered Ctenophora-sister with only 69% support. This study did not include Trichoplax.

Philippe et al. 2009

Philippe et al. 2009 [Philippe:2009hh] assembled an EST dataset for 55 species with 128 genes to explore phylogenetic relationship of early diverging animals. The phylogenetic analysis using Poisson-CAT model strongly supported Porifera-sister, and ctenophores was sister to cnidarians to form the “coelenterate” clade. The authors concluded: > The resulting phylogeny yields two significant conclusions reviving old views that have been challenged in the molecular era: (1) that the sponges (Porifera) are monophyletic and not paraphyletic as repeatedly proposed, thus undermining the idea that ancestral metazoans had a sponge-like body plan; (2) that the most likely position for the ctenophores is together with the cnidarians in a ‘‘coelenterate’’ clade.

Hejnol et al. 2009

Hejnol et al. 2009 added EST sequences from seven taxa, and a total of 94 taxa were included in the final data matrix. The orthology inference was largely simialr to Dunn et al. 2008. Two datsets were constructed, one with 844 genes and the other with 330 genes. Maximum likelihood analyses from both datasets strongly supported Ctenophora-sister.

Pick et al. 2010

Pick et al.³⁸ sought to test whether Ctenophora-sister was an artefact of insufficient taxon sampling. They added new and additional published sequence data to the 64-taxon matrix of Dunn et al.². The new taxa included 12 sponges, 1 ctenophore, 5 cnidarians, and Trichoplax. They further modified the matrix by removing 2,150 sites that were poorly sampled or aligned. They considered two different sets of outgroups: Choanoflagellata (resulting in Choanimalia) and the same sampling as Dunn et al. (resulting in Opisthokonta).

All their analyses were conducted under the F81+CAT+Gamma model in phylobayes³, in both a Bayesian framework and with bootstrapping. All analyses have the same ingroup sampling and site removal so it isn’t possible to independently assess the impact of these factors. Analyses with Choanimalia sampling recovered Porifera-sister with 72% posterior probability (PP) and 91% bootstrap support (BS). With broader Opisthokonta sampling, support for Porifera-sister is 84% PP. This is an interesting case where increased outgroup sampling leads to increased support for Porifera-sister.

The authors argue that previous results supporting Ctenophora-sister “are artifacts stemming from insufficient taxon sampling and long-branch attraction (LBA)” and that “this hypothesis should be rejected”. Although the posterior probabilities supporting Porifera-sister are not strong, they conclude:

Results of our analyses indicate that sponges are the sister group to the remaining Metazoa, and Placozoa are sister group to the Bilateria

They also investigated saturation, and conclude that Dunn et al.² is more saturated than Philippe et al. 2009 [Philippe:2009hh]. Note that the Pick et al.³⁸ dataset is not reanalyzed here because partition data are not available, and due to site filtering the partition file from Dunn et al.² cannot be applied to this matrix.

Ryan et al. 2013

Ryan et al. 2013 sequenced the first ctenophore genome of Mnemiopsis leidyi. With the genome resources of M. leidyi, the authors constructed two phylogenomic datasets: a “Genome set” based on 13 animal genomes and a “EST Set” that also included 59 animals. They analyzed both matrices by site-homogeneous GTR+Gamma model and Poisson-CAT model with three sets of outgroup sampling to evaluate the effect of outgroup selection to the ingroup topology. The results strongly supported Ctenophora-sister in all datasets they analysed using site-homogeneous model. The Poisson + CAT model of the genome dataset strongly supported of a clade of Ctenophora and Porifera as the sister group to all other Metazoa and Bayesian analysis on the EST dataset did not converge after 205 days.

Moroz et al. 2014

Moroz et al. 2014 sequenced the second ctenophore genome Pleurobrachia bachei to explore the phylogenetic relationship of Metazoa. All phylogenetic analyses strongly supported Ctenophora-sister with different taxon and gene sampling using WAG site-homogeneous model.

Whelan et al. 2015

Wehlan et al. 2015 constructed a new phylogenomic dataset by eight new transcripomic data and investigated a range of possible sources of systematic error under multiple analyses (e.g. long-branch attraction, compositional bias, fast evolving genes, etc.). Putative orthologs were determined of each species using HaMStR using the model organism core ortholog set (~ 1000 orthologs) and subsequetly removal of genes with too much missing data and potential paralogs. The authors further filtered the full dataset to 24 sub-datasets by filtering genes with high long-branch scores; genes with high RSFV values; genes that are potential paralogs; fast evolving genes and progressively removal of outgroups. All the maximum likelihood analyses with site-homogeneous model and PartitionFinder strongly suggested Ctenophora-sister. CAT-GTR models only used in least saturated dataset 6 and 16 also strongly supported Ctenophora. Regarding the recovery of Ctenophora-sister, the authors concluded: > Importantly, biases resulting from elevated compositional heterogeneity or elevated substitution rates are ruled out. Placement of ctenophores as sister to all other animals, and sponge monophyly, are strongly supported under multiple analyses, herein.

Chang et al. 2015

Chang et al. 2015 was originally used to explore phylogenetic position of Myxozoa in Cnidaria but also sampled broadly across the breadth of animal diversity. The authors constructed a dataset with 200 protein markers based on Philippe et al. 2011 with 51,940 amino acids and 77 taxa. Both site-heterogeneous Poisson-CAT and site-homogeneous GTR models strongly supported Ctenophora-sister.

Pisani et al. 2015

Pisani et al. 2015 reanalyzed representative datasets that supported Ctenophora-sister, including Ryan2013, Moroz2014 and Whelan2015 datasets. It was the first study showing that progressively removal of more distantly related outgroups could largely affect phylogenomic inference of animal-root position. The authors suggested that the inclusion of outgroups very distant from the ingroup can cause systematic errors due to long-branch attraction. They found strong support of porifera-sister when only include Choanozoa as outgroup in these datasets using site-heterogeous CAT model. Regarding the recovery of Porifera-sister, the authors concluded: >Our results reinforce a traditional scenario for the evolution of complexity in animals, and indicate that inferences about the evolution of Metazoa based on the Ctenophora-sister hypothesis are not supported by the currently available data.

Feuda et al. 2017

Feuda et al. didn’t generate any new data, instead they used the data-recoding methods to reanalyze two key datasets that support Ctenophora-sister (Whelan2015_D20, Chang2015 datasets). It was the first study phylogenomic analysis using recoding strategy to explore animal-root position. The authors compared model adequacy using posterior predictive analyses from a set of site-homogeneous (WAG, LG, GTR, PartitionFinder)and site-heterogeneous (CAT-GTR) models with non-recoding and recoding datasets. The results showed that data-recoding can significant reduce compositional heterogeneity in both datasets with CAT-GTR models and strongly supported Porifera-sister hypothesis. Regarding the recovery of Porifera-sister, the authors concluded:

Because adequate modeling of the evolutionary process that generated the data is fundamental to recovering an accurate phylogeny, our results strongly support sponges as the sister group of all other animals and provide further evidence that Ctenophora-sister represents a tree reconstruction artifact.

Whelan and Halalnych 2017a

Whelan et al. is the only study to evaluate performance of site-heterogeneous models and site-homogeneous model with data partitioning under the simulation framework. The simulation results suggested that Poisson+CAT model consistently performed worse than other models in simulation datasets. More importantly, the authors also showed that both Poisson + CAT and GTR + CAT models could overestimated substitutional heterogeneity in many cases. They also reanalyzed datasets from Philippe 2009 and Nosenko 2013 using both CAT models and data partitioning with site -homogeneious model. The results indicated that Poisson + CAT model tends to recover less accurate trees and both GTR + CAT and data partitioning strongly supported Ctenophora-sister in reanalyses. The authors also concluded:

Practices such as removing constant sites and parsimony uninformative characters, or using CAT-F81 when CAT-GTR is deemed too computationally expensive, cannot be logically justified. Given clear problems with CAT-F81, phylogenies previously inferred with this model should be reassessed.

Whelan et al. 2017b

Whelan et al. added 27 new ctenophore transcriptomic data to explore animal-root position as well as relationships within Ctenophores. It significantly increased ctenophore taxon sampling than other studies. Putative orthologs were determined largely similar to Whelan2015, with the difference of a core Ctenophora core datasets were constructed here with more than 2000 genes. The subsequent filtering strategy was also similar to the previous study. All analyses using site-homogeneous and site-heterogeneous models strongly supported Ctenophora-sister hypothesis, even with CAT-GTR model in Choanoazoa dataset. Regarding the recovery of Ctenophora-sister, the authors concluded: >Using datasets with reasonably high ctenophore and other non-bilaterian taxon sampling, our results strongly reject the hypothesis that sponges are the sister lineage to all other extant metazoans.

Simion et al. 2017

Simion et al.² added transcriptomic data for 21 new animals. The data matrix was constructed using a semi-automated approach to comprehensively detect and eliminate potential systematic errors. The resulting dataset comprises 1,719 genes and 97 species, including 61 non-bilaterian metazoan species. It was by far the largest phylogenomic dataset in terms of taxon and gene sampling related to animal-root question.

The supermatrix was first analyzed using the Poisson-CAT model. Different that other phylobayes analyses, Simion et al. used a gene jackknife strategy based on 100 analyses to overcome the computational limitation because of the large data size. Each jackknife is baed on a random selection of ~ 25% of the genes. The Phylobayes with site-heterogeneous model strongly supported the Porifera-sister, whereas site-homogeneous strongly supported Ctenophora-sister in the supermatrix dataset. Importantly, the authors compared the behavior of long-branch effect between site-heterogenious and site-homogeneous models by progressively removal taxa and concluded higher sensitivity of site-homogeneous models to LBA than CAT models. Regarding the recovery of Ctenophora-sister, the authors concluded:

Our dataset outperforms previous metazoan gene superalignments in terms of data quality and quantity. Analyses with a best-fitting site-heterogeneous evolutionary model provide strong statistical support for placing sponges as the sister-group to all other metazoans, with ctenophores emerging as the second-earliest branching animal lineage.

Taxa comparison across matrices

The primary intent of comparing taxa across matrices was to validate our taxon name reconciliation across studies.

We first considered pairwise similarity between the same species from different matrices in different studies.

Partition comparison across matrices

The count for a partition pair can be much larger than the number of genes in the matrix, which suggests that the count is the number of hsps rather than the number of sequences with hits.

There are 17 matrices. A gene that is perfectly sampled would form a cluster with this size. To our surprise, very few clusters, though, are this size. This suggests that intersection of genes between matrices is extreme low. This finding suggested genes used for exploring the position of the root of the animal phylogeny